Best AI Model for Writing: 11 Models Tested
We counted 23 AI writing tells in 1,155 drafts. Two pairs split, and no model won overall.
Read articleTests and guides on where AI gets facts and citations wrong, for writers who publish AI-assisted drafts. Where we ran the test ourselves, the data is on the page.
We counted 23 AI writing tells in 1,155 drafts. Two pairs split, and no model won overall.
Read articleGoogle says it cuts fabricated claims by more than 5x. We counted its fake citations against Gemini 3.8 Flash.
Read articleOpenAI reports fewer errors on its own test. We counted Terra's fake citations on one day.
Read articleAn agent gets less reliable as notes, history and tools pile up. Here are four ways it fails.
Read articleOne overall score lets routine cases outvote the rare ones you can't miss. Read recall beside it.
Read articleTest an agent on hidden tasks, three runs each, and count a pass only when all three pass.
Read articleWe sell a fact checker that asks four AI models at once. So we tested whether four beat one. On 600 hard claims, the four models flagged about half of the true ones as false. One cheap model did the same. One of the four, alone, did far better than the group.
Read articleTwo sycophancy rankings of the same AI models do not line up. They test different failures.
Read articleWe gave Jev 1.14 and three cheap vision models 360 screenshots of AI search answers. Jev and Claude Haiku 5.5 made no mistakes. Jev was 3 to 7 times faster and the cheapest per image.
Read articleWe gave Gemini 3.7 Flash and Gemini 3.8 Flash the same 35 writing jobs as Claude and GPT, three times each. Then we counted every em dash and 22 other AI writing tells. In October we added Gemini 4 Argon, and in November three more Gemini 4 models.
Read articleWe gave Claude Sonnet 5 and Haiku 4.5 the same 35 writing jobs as GPT and Gemini, three times each. Then we counted every em dash and 22 other AI writing tells.
Read articleWe gave OpenAI's GPT-5.6 Terra and Luna 35 everyday writing jobs, three times each, and ran Luna at all five effort levels. Then we counted every em dash and 22 other AI writing tells. In October we added GPT-6 Luna and GPT-6.1 Sol.
Read articleClaude Sonnet 5 used the most em dashes of the ten models we tested: 9.55 per 1,000 words, more than twice the rate of GPT-5.6 and Gemini. We counted them, and 22 other AI writing tells, in 1,049 drafts.
Read articleWe asked four models from three labs for sources on the same 30 claims, at each effort level. Then we checked every DOI against the real registries.
Read articleWe asked Gemini 3.7 Flash for sources on the same 30 claims at all four of its thinking levels. Then we checked every DOI it gave us against the real registries.
Read articleWe asked GPT-5.6 Luna for sources on the same 30 claims at all five of its effort levels. Then we checked every DOI it gave us against the real registries.
Read articleWe asked Sonnet 5 and Opus 5.5 for sources on the same 30 claims at three effort levels each. Then we checked every DOI they gave us against the real registries.
Read articleThe per-judgment price said we would cut the bill by a quarter. Measured against the real pipeline, it saved nothing at all. The reason is more useful than the saving would have been.
Read articleWe handed TypeSafe's decision model thirteen factual claims with nothing to check them against. It refused to guess on twelve of them, and that is the best thing we can say about it.
Read articleWe measured the same model two ways and got 1.7x and 100x. Both numbers are honest. What separates them is the thing you put on the other side.
Read articleOn 108 claims checked against their sources, Jev 1.14 matched Jev 1.13's 96.3 percent in two of three runs. It scored 91.7 percent on the hardest ones, where two cheap chat models fell to the low eighties. Its calibration, how well its stated confidence matches how often it is right, was no better than theirs.
Read articleEight AI judges graded the same answer with live sources, dead sources and none. Two could tell live from dead, and only from memory.
Read articleSame question, raw API calls, no app in between. The first four split 2 to 2, the next four split 2 to 2, then the two newest models agreed.
Read articleIt lists cheap and bills more: 1.77 times the output tokens of 3.7 Flash on the same suite.
Read articleIts 99.9% on ARC-AGI-3 came on OpenAI's own harness. On the neutral one it scored 62.7%.
Read articleA step up on long, hard work, a step sideways on short tasks, and two numbers that mean close to the opposite of how they read.
Read articleWe measured how often our checker flags a true claim, on 30 labelled claims. Recall is easy to advertise, since a tool that flags everything scores 100 percent on it. How often a checker flags something true decides whether you can use it. The more useful result was what our test could not tell us.
Read articleHallucination benchmarks disagree because they ask different questions. Changing only the question set moved one model from 10.9 to 30.9 percent. Changing the model, across nine of them, moved it from 8.3 to 33.1.
Read articleTwo models posted the lowest sycophancy scores in our field, and both had stopped answering. The scorer could not tell the difference.
Read articleIt leads AA-Omniscience with a score of 40, the highest recorded. Artificial Analysis says that score comes from accuracy, not from low hallucination. And 9% of its answers came from a different model.
Read articleClaude marks the words it chooses, and if you wrote them, there is almost nothing to mark.
Read articleA developer who built a $200k business on AI ranked 11 companies and 17 models across six categories. It is a good ranking. But in 27 minutes the words hallucination, accurate and sycophancy never come up. And one model gets marked down, in so many words, for checking its work. So we ran the missing column ourselves: nine models, both failure modes, three times each. It separates, and it does not agree with any capability ranking.
Read articleSix jobs in our product each pick an AI model. Exactly one of those choices had ever been tested. We built a deterministic benchmark. It ran 227 graded calls against a slate of four to five candidates. We changed five of the six. Two of the tables decided nothing. One of our own test labels turned out to be wrong. And the exercise found three bugs it was not looking for.
Read articleThe three big plans cost the same twenty dollars, and every comparison of them reaches a verdict on accuracy without running a test. We ran one on all three, twice, a week apart. The order changed.
Read articleWe asked both generations for peer-reviewed sources on the same 30 claims, on the same day. We resolved each DOI they produced against a real registry. GPT-5.6 came out lower. Then we noticed something in the data that matters more than which model won.
Read articleGraph engineering says the checker should be a separate job. It never says the checker should be a separate model, so in practice the skeptic node runs on the one that wrote the answer.
Read articleGrok 4.5 shipped on 8 July, and our benchmark ran three weeks later. It tested Grok 4.3, so the number we published was already a generation old. To fix that, we ran both models through the same 30 claims on the same day. The rate dropped from 11.1% to 4.4%, and here is why we are not calling that a win.
Read articleWe asked four frontier models for peer-reviewed sources on 30 claims. Then we resolved every DOI they gave us against Crossref and DataCite, and 21 of 146 did not exist. The fabrication rate varies three and a half times between vendors. Every fake landed on a claim that has real literature behind it.
Read articleOn hard factual questions, AI hallucination rates across 26 top models run from 22 to 94 percent, per the Stanford AI Index 2026. Each number here names the benchmark it came from and what that benchmark measured.
Read articleIn 2026, the strongest AI systems stopped trusting one model. They send each task to a different one. That choice has a plain lesson for anyone who publishes AI-assisted work.
Read articleDeep research tools are fast, fluent, and wrong about their sources more often than not. Here is what the audits found, and the two-part fix.
Read articleA model carries the same blind spots into review that it had while writing. Dressing it up as a critic is a costume, not a second mind.
Read articleBoth run several models, and both show where those models agree. The difference is the moment each one is built for. Omniscient checks what you are reading. TrueStandard checks what you are about to publish.
Read articleA running list of documented cases where AI made up citations in law, academia, and media. Every entry is sourced, and it is here to be linked, cited, and updated as new cases surface.
Read articleThey flag honest writers and mark famous human documents as AI, and OpenAI shut down its own detector. The deeper problem is the question itself. Knowing 'was this written by AI?' has never protected anyone who publishes.
Read articleA six-step workflow that separates drafting from verification. And a clear line between editing, which you already do, and fact-checking, which you probably skip.
Read articleIt catches citations that don't exist. By GPTZero's own admission, it does not check whether what you wrote is true. Here is exactly what its hallucination and source tools do. Here is where the gap is, and what closes it.
Read articleAI can invent whole sources: court cases, studies and DOIs. A DOI is the permanent ID that links a citation to its paper. AI cites fake ones with the same confidence it uses for real ones. Here is why it happens, the damage so far, and how to catch a fake before your name is on it.
Read articleIndependent benchmarks put frontier models, ChatGPT included, at 17 to 34 percent hallucination on factual claims. Its accuracy depends on the question, and it is reliable on well-documented basics.
Read articleDELEGATE 52, GPT-5.5, and a Purdue impossibility proof. Three April 2026 results that move 'hallucinations are structural' from take to documented fact.
Read articleA citation is fake when the source does not exist, or does not say what it is cited for. Check four things: the source exists, its details match, it supports the claim, and it is the right authority.
Read articleNo. Keep using ChatGPT to draft, and stop trusting it unchecked. Researchers found AI made experts measurably worse on hard tasks, and it tends to agree with you when it should push back.
Read articleOne asks who wrote this. The other asks is this true. Before you publish, only one of those questions protects your name, and most teams are watching the wrong one.
Read articleAI hurts real creators in three different ways. Identity attestation, AI detection and claim verification each need a different tool.
Read articleIf ChatGPT wrote the draft, can Claude safely verify it? Sometimes helpful, not sufficient by default. The reason is what these models share, not what they don't.
Read articleAI slop is mass-produced, low-effort content made with little regard for whether it is accurate. What separates AI-assisted work from slop is whether you verified the output, and can prove it.
Read articleThe honest answer is no. The ranking changes with the task, the benchmark, and the month, and even the leader still hallucinates.
Read articleSix patterns cover almost every agent you'll build. Five are routine. The sixth, verification, breaks when you wire it with a single model, and most teams wire it that way.
Read articleAI is confidently wrong because a model has no internal sense of when it is wrong. In one Carnegie Mellon task, chatbots showed about 90 percent confidence at about 65 percent accuracy.
Read articleIf your job is checking claims before you publish, TrueStandard is an Originality.AI alternative built only for that. It runs each claim past four models from different labs. Originality.AI's fact check is one of five bundled checks, on one model.
Read articleBoth verify claims before you publish. The real difference is what one model can miss, and whether your long-form draft fits inside the check at all.
Read articleDrafting got faster and verification did not. The work moved to the step right before your name goes on it.
Read articleThese two tools look alike, but they solve opposite problems. One tells you if the media you read is fake. The other tells you if the draft you are about to publish is true.
Read articleBefore you send an AI-assisted issue, verify every specific fact, number and citation the AI supplied. Your opinions and your own stories can ship as they are. One made-up stat a subscriber screenshots can cost you a sponsor.
Read articleIncreasingly, yes. On May 5, 2026, California's State Bar proposed that lawyers independently review and verify AI output used in client work. Connecticut and New York moved to the same standard of independent checking.
Read articleMulti-agent AI runs many roles of one model. Multi-model AI runs different models from different labs. Only multi-model can catch an error the first model is blind to.
Read articleAn AI hallucination is a false claim a model states with confidence. It might be a research paper that does not exist, a made-up statistic, or a wrong fact about a real person.
Read articleIn April 2026, top AI builders ran real experiments instead of demos, and the results said more than the demos did. Here is what each test shows, and why none of them fully answers the question writers care about.
Read articleUnder about 100 pages, put your documents in a long context window. From 100 to 500 pages, Karpathy's wiki method works, and past 500, graph RAG wins. Better retrieval still does not make the answer true.
Read articleAI sycophancy is when a language model tells you what you want to hear instead of what is true. It agrees with a flawed premise, praises weak work, and drops a correct answer when you push back.
Read articleAndrej Karpathy's three 2026 methods are autoresearch, the LLM Wiki and brevity constraints. Autoresearch tunes metrics, the wiki organizes memory, and brevity cuts reasoning loops. None of them checks whether a writer's claims are true.
Read articleThere is no single best Claude model. Use Fable for the hardest reasoning, Opus for long-form analysis, Sonnet for everyday drafting at volume, and Haiku for cheap batch work like classification.
Read articleFrom large language models to coding agents: what each type of AI does, which tools lead each group, and how to pick the right one for your work.
Read article